kube-state-metrics (KSM) is a simple service that listens to the Kubernetes API server and generates metrics about the state of the objects. (See examples in the Metrics section below.) It is not focused on the health of the individual Kubernetes components, but rather on the health of the various objects inside, such as deployments, nodes and pods.
kube-state-metrics is about generating metrics from Kubernetes API objects without modification. This ensures that features provided by kube-state-metrics have the same grade of stability as the Kubernetes API objects themselves. In turn, this means that kube-state-metrics in certain situations may not show the exact same values as kubectl, as kubectl applies certain heuristics to display comprehensible messages. kube-state-metrics exposes raw data unmodified from the Kubernetes API, this way users have all the data they require and perform heuristics as they see fit.
The metrics are exported on the HTTP endpoint /metrics on the listening port
(default 8080). They are served as plaintext. They are designed to be consumed
either by Prometheus itself or by a scraper that is compatible with scraping a
Prometheus client endpoint. You can also open /metrics in a browser to see
the raw metrics. Note that the metrics exposed on the /metrics endpoint
reflect the current state of the Kubernetes cluster. When Kubernetes objects
are deleted they are no longer visible on the /metrics endpoint.
Note
This README is generated from a template. Please make your changes there and run make generate-template.
- Versioning
- Metrics Documentation
- Kube-state-metrics self metrics
- Resource recommendation
- Latency
- A note on costing
- kube-state-metrics vs. metrics-server
- Scaling kube-state-metrics
- Setup
- Usage
kube-state-metrics uses client-go to talk with
Kubernetes clusters. The supported Kubernetes cluster version is determined by client-go.
The compatibility matrix for client-go and Kubernetes cluster can be found
here.
All additional compatibility is only best effort, or happens to still/already be supported.
At most, 5 kube-state-metrics and 5 kubernetes releases will be recorded below. Generally, it is recommended to use the latest release of kube-state-metrics. If you run a very recent version of Kubernetes, you might want to use an unreleased version to have the full range of supported resources. If you run an older version of Kubernetes, you might need to run an older version in order to have full support for all resources. Be aware, that the maintainers will only support the latest release. Older versions might be supported by interested users of the community.
| kube-state-metrics | Kubernetes client-go Version |
|---|---|
| v2.10.1 | v1.27 |
| v2.11.0 | v1.28 |
| v2.12.0 | v1.29 |
| v2.13.0 | v1.30 |
| v2.14.0 | v1.31 |
| main | v1.31 |
Resources in Kubernetes can evolve, i.e., the group version for a resource may change from alpha to beta and finally GA in different Kubernetes versions. For now, kube-state-metrics will only use the oldest API available in the latest release.
The latest container image can be found at:
registry.k8s.io/kube-state-metrics/kube-state-metrics:v2.14.0(arch:amd64,arm,arm64,ppc64leands390x)- View all multi-architecture images at here
Any resources and metrics based on alpha Kubernetes APIs are excluded from any stability guarantee, which may be changed at any given release.
See the docs directory for more information on the exposed metrics.
The *_labels family of metrics exposes Kubernetes labels as Prometheus labels.
As Kubernetes
is more liberal than
Prometheus
in terms of allowed characters in label names,
we automatically convert unsupported characters to underscores.
For example, app.kubernetes.io/name becomes label_app_kubernetes_io_name.
This conversion can create conflicts when multiple Kubernetes labels like
foo-bar and foo_bar would be converted to the same Prometheus label label_foo_bar.
Kube-state-metrics automatically adds a suffix _conflictN to resolve this conflict,
so it converts the above labels to
label_foo_bar_conflict1 and label_foo_bar_conflict2.
If you'd like to have more control over how this conflict is resolved, you might want to consider addressing this issue on a different level of the stack, e.g. by standardizing Kubernetes labels using an Admission Webhook that ensures that there are no possible conflicts.
kube-state-metrics exposes its own general process metrics under --telemetry-host and --telemetry-port (default 8081).
kube-state-metrics also exposes list and watch success and error metrics. These can be used to calculate the error rate of list or watch resources. If you encounter those errors in the metrics, it is most likely a configuration or permission issue, and the next thing to investigate would be looking at the logs of kube-state-metrics.
Example of the above mentioned metrics:
kube_state_metrics_list_total{resource="*v1.Node",result="success"} 1
kube_state_metrics_list_total{resource="*v1.Node",result="error"} 52
kube_state_metrics_watch_total{resource="*v1beta1.Ingress",result="success"} 1
kube-state-metrics also exposes some http request metrics, examples of those are:
http_request_duration_seconds_bucket{handler="metrics",method="get",le="2.5"} 30
http_request_duration_seconds_bucket{handler="metrics",method="get",le="5"} 30
http_request_duration_seconds_bucket{handler="metrics",method="get",le="10"} 30
http_request_duration_seconds_bucket{handler="metrics",method="get",le="+Inf"} 30
http_request_duration_seconds_sum{handler="metrics",method="get"} 0.021113919999999998
http_request_duration_seconds_count{handler="metrics",method="get"} 30
kube-state-metrics also exposes build and configuration metrics:
kube_state_metrics_build_info{branch="main",goversion="go1.15.3",revision="6c9d775d",version="v2.0.0-beta"} 1
kube_state_metrics_shard_ordinal{shard_ordinal="0"} 0
kube_state_metrics_total_shards 1
kube_state_metrics_build_info is used to expose version and other build information. For more usage about the info pattern,
please check the blog post here.
Sharding metrics expose --shard and --total-shards flags and can be used to validate
run-time configuration, see /examples/prometheus-alerting-rules.
kube-state-metrics also exposes metrics about it config file and the Custom Resource State config file:
kube_state_metrics_config_hash{filename="crs.yml",type="customresourceconfig"} 2.38272279311849e+14
kube_state_metrics_config_hash{filename="config.yml",type="config"} 2.65285922340846e+14
kube_state_metrics_last_config_reload_success_timestamp_seconds{filename="crs.yml",type="customresourceconfig"} 1.6704882592037103e+09
kube_state_metrics_last_config_reload_success_timestamp_seconds{filename="config.yml",type="config"} 1.6704882592035313e+09
kube_state_metrics_last_config_reload_successful{filename="crs.yml",type="customresourceconfig"} 1
kube_state_metrics_last_config_reload_successful{filename="config.yml",type="config"} 1
Resource usage for kube-state-metrics changes with the Kubernetes objects (Pods/Nodes/Deployments/Secrets etc.) size of the cluster. To some extent, the Kubernetes objects in a cluster are in direct proportion to the node number of the cluster.
As a general rule, you should allocate:
- 250MiB memory
- 0.1 cores
Note that if CPU limits are set too low, kube-state-metrics' internal queues will not be able to be worked off quickly enough, resulting in increased memory consumption as the queue length grows. If you experience problems resulting from high memory allocation or CPU throttling, try increasing the CPU limits.
In a 100 node cluster scaling test the latency numbers were as follows:
"Perc50": 259615384 ns,
"Perc90": 475000000 ns,
"Perc99": 906666666 ns.
By default, kube-state-metrics exposes several metrics for events across your cluster. If you have a large number of frequently-updating resources on your cluster, you may find that a lot of data is ingested into these metrics. This can incur high costs on some cloud providers. Please take a moment to configure what metrics you'd like to expose, as well as consult the documentation for your Kubernetes environment in order to avoid unexpectedly high costs.
The metrics-server is a project that has been inspired by Heapster and is implemented to serve the goals of core metrics pipelines in Kubernetes monitoring architecture. It is a cluster level component which periodically scrapes metrics from all Kubernetes nodes served by Kubelet through Metrics API. The metrics are aggregated, stored in memory and served in Metrics API format. The metrics-server stores the latest values only and is not responsible for forwarding metrics to third-party destinations.
kube-state-metrics is focused on generating completely new metrics from Kubernetes' object state (e.g. metrics based on deployments, replica sets, etc.). It holds an entire snapshot of Kubernetes state in memory and continuously generates new metrics based off of it. And just like the metrics-server it too is not responsible for exporting its metrics anywhere.
Having kube-state-metrics as a separate project also enables access to these metrics from monitoring systems such as Prometheus.
In order to shard kube-state-metrics horizontally, some automated sharding capabilities have been implemented. It is configured with the following flags:
--shard(zero indexed)--total-shards
Sharding is done by taking an md5 sum of the Kubernetes Object's UID and performing a modulo operation on it with the total number of shards. Each shard decides whether the object is handled by the respective instance of kube-state-metrics or not. Note that this means all instances of kube-state-metrics, even if sharded, will have the network traffic and the resource consumption for unmarshaling objects for all objects, not just the ones they are responsible for. To optimize this further, the Kubernetes API would need to support sharded list/watch capabilities. In the optimal case, memory consumption for each shard will be 1/n compared to an unsharded setup. Typically, kube-state-metrics needs to be memory and latency optimized in order for it to return its metrics rather quickly to Prometheus. One way to reduce the latency between kube-state-metrics and the kube-apiserver is to run KSM with the --use-apiserver-cache flag. In addition to reducing the latency, this option will also lead to a reduction in the load on etcd.
Sharding should be used carefully and additional monitoring should be set up in order to ensure that sharding is set up and functioning as expected (eg. instances for each shard out of the total shards are configured).
Automatic sharding allows each shard to discover its nominal position when deployed in a StatefulSet which is useful for automatically configuring sharding. This is an experimental feature and may be broken or removed without notice.
To enable automated sharding, kube-state-metrics must be run by a StatefulSet and the pod name and namespace must be handed to the kube-state-metrics process via the --pod and --pod-namespace flags. Example manifests demonstrating the autosharding functionality can be found in /examples/autosharding.
This way of deploying shards is useful when you want to manage KSM shards through a single Kubernetes resource (a single StatefulSet in this case) instead of having one Deployment per shard. The advantage can be especially significant when deploying a high number of shards.
The downside of using an auto-sharded setup comes from the rollout strategy supported by StatefulSets. When managed by a StatefulSet, pods are replaced one at a time with each pod first getting terminated and then recreated. Besides such rollouts being slower, they will also lead to short downtime for each shard. If a Prometheus scrape happens during a rollout, it can miss some of the metrics exported by kube-state-metrics.
For pod metrics, they can be sharded per node with the following flag:
--node=$(NODE_NAME)
Each kube-state-metrics pod uses FieldSelector (spec.nodeName) to watch/list pod metrics only on the same node.
A daemonset kube-state-metrics example:
apiVersion: apps/v1
kind: DaemonSet
spec:
template:
spec:
containers:
- image: registry.k8s.io/kube-state-metrics/kube-state-metrics:IMAGE_TAG
name: kube-state-metrics
args:
- --resource=pods
- --node=$(NODE_NAME)
env:
- name: NODE_NAME
valueFrom:
fieldRef:
apiVersion: v1
fieldPath: spec.nodeName
To track metrics for unassigned pods, you need to add an additional deployment and set --track-unscheduled-pods, as shown in the following example:
apiVersion: apps/v1
kind: Deployment
spec:
template:
spec:
containers:
- image: registry.k8s.io/kube-state-metrics/kube-state-metrics:IMAGE_TAG
name: kube-state-metrics
args:
- --resources=pods
- --track-unscheduled-pods
Other metrics can be sharded via Horizontal sharding.
Install this project to your $GOPATH using go get:
go get k8s.io/kube-state-metrics
Simply run the following command in this root folder, which will create a self-contained, statically-linked binary and build a Docker image:
make container
Simply build and run kube-state-metrics inside a Kubernetes pod which has a service account token that has read-only access to the Kubernetes cluster.
The (kube-prometheus) stack installs kube-state-metrics as one of its components; you do not need to install kube-state-metrics if you're using the kube-prometheus stack.
If you want to revise the default configuration for kube-prometheus, for example to enable non-default metrics, have a look at Customizing Kube-Prometheus.
To deploy this project, you can simply run kubectl apply -f examples/standard and a Kubernetes service and deployment will be created. (Note: Adjust the apiVersion of some resource if your kubernetes cluster's version is not 1.8+, check the yaml file for more information).
To have Prometheus discover kube-state-metrics instances it is advised to create a specific Prometheus scrape config for kube-state-metrics that picks up both metrics endpoints. Annotation based discovery is discouraged as only one of the endpoints would be able to be selected, plus kube-state-metrics in most cases has special authentication and authorization requirements as it essentially grants read access through the metrics endpoint to most information available to it.
Note: Google Kubernetes Engine (GKE) Users - GKE has strict role permissions that will prevent the kube-state-metrics roles and role bindings from being created. To work around this, you can give your GCP identity the cluster-admin role by running the following one-liner:
kubectl create clusterrolebinding cluster-admin-binding --clusterrole=cluster-admin --user=$(gcloud info --format='value(config.account)')
Note that your GCP identity is case sensitive but gcloud info as of Google Cloud SDK 221.0.0 is not. This means that if your IAM member contains capital letters, the above one-liner may not work for you. If you have 403 forbidden responses after running the above command and kubectl apply -f examples/standard, check the IAM member associated with your account at https://console.cloud.google.com/iam-admin/iam?project=PROJECT_ID. If it contains capital letters, you may need to set the --user flag in the command above to the case-sensitive role listed at https://console.cloud.google.com/iam-admin/iam?project=PROJECT_ID.
After running the above, if you see Clusterrolebinding "cluster-admin-binding" created, then you are able to continue with the setup of this service.
The following healthcheck endpoints are available (self refers to the telemetry port, while main refers to the exposition port):
/healthz(exposed onmain): Returns a 200 status code if the application is running. We recommend to use this for the startup probe./livez(exposed onmain): Returns a 200 status code if the application is not affected by an outage of the Kubernetes API Server. We recommend to using this for the liveness probe./readyz(exposed onself): Returns a 200 status code if the application is ready to accept requests and expose metrics. We recommend using this for the readiness probe.
Note that it is discouraged to use the telemetry metrics endpoint for any probe when proxying the exposition data.
If you want to run kube-state-metrics in an environment where you don't have cluster-reader role, you can:
- create a serviceaccount
apiVersion: v1
kind: ServiceAccount
metadata:
name: kube-state-metrics
namespace: your-namespace-where-kube-state-metrics-will-deployed- give it
viewprivileges on specific namespaces (using roleBinding) (note: you can add this roleBinding to all the NS you want your serviceaccount to access)
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
name: kube-state-metrics
namespace: project1
roleRef:
apiGroup: rbac.authorization.k8s.io
kind: ClusterRole
name: view
subjects:
- kind: ServiceAccount
name: kube-state-metrics
namespace: your-namespace-where-kube-state-metrics-will-deployed- then specify a set of namespaces (using the
--namespacesoption) and a set of kubernetes objects (using the--resources) that your serviceaccount has access to in thekube-state-metricsdeployment configuration
spec:
template:
spec:
containers:
- name: kube-state-metrics
args:
- '--resources=pods'
- '--namespaces=project1'For the full list of arguments available, see the documentation in docs/developer/cli-arguments.md
Starting from the kube-state-metrics chart v2.13.3 (kube-state-metrics image v1.9.8), the official Helm chart is maintained in prometheus-community/helm-charts. Starting from kube-state-metrics chart v3.0.0 only kube-state-metrics images of v2.0.0 + are supported.
When developing, test a metric dump against your local Kubernetes cluster by running:
Users can override the apiserver address in KUBE-CONFIG file with
--apiservercommand line.
go install
kube-state-metrics --port=8080 --telemetry-port=8081 --kubeconfig=<KUBE-CONFIG> --apiserver=<APISERVER>
Then curl the metrics endpoint
curl localhost:8080/metrics
To run the e2e tests locally see the documentation in tests/README.md.
When developing, there are certain code patterns to follow to better your contributing experience and likelihood of e2e and other ci tests to pass. To learn more about them, see the documentation in docs/developer/guide.md.
This project is sponsored by SIG Instrumentation.
There is also a channel for #kube-state-metrics on Kubernetes' Slack.
You can also join the SIG Instrumentation mailing list. This will typically add invites for the following meetings to your calendar, in which topics around kube-state-metrics can be discussed.
- Regular SIG Meeting: Thursdays at 9:30 PT (Pacific Time) (biweekly). Convert to your timezone.
- Regular Triage Meeting: Thursdays at 9:30 PT (Pacific Time) (biweekly - alternating with regular meeting). Convert to your timezone.